GeneSys: Generative Modeling of Developmental System

Abstract

Temporal single-cell transcriptomics enables the reconstruction of dynamic gene expression changes during development. Yet, its analytical power is often limited by data sparsity, technical noise, and imbalanced representation of cell types across time points. To overcome these challenges, we present GeneSys (Generative Modeling of Developmental System), a generative deep learning model that simulates single-cell transcriptomic landscapes under developmental constraints, which is informed by prior biological knowledge or user-defined hypotheses. GeneSys integrates a temporal variational autoencoder with a cell-type classifier, requiring a lineage blueprint as input, which enables it to model the temporal transitions of transcriptional states with cell-type specificity. Leveraging data from Arabidopsis thaliana roots and mouse embryos, we show that GeneSys learns robust developmental trajectories, generates imputed and representative transcriptomes, and enhances gene prioritization accuracy compared to unregularized scRNA-seq data. By applying gene masking and augmentation, GeneSys reveals interpretable gene expression programs (GEPs) and serves as an in silico platform to test the impact of specific genes or gene sets on user-defined developmental outcomes. Additionally, GeneSys computes linear interaction matrices (LIMAs) to infer dynamic gene networks and prioritize transcription factors with spatiotemporal resolution. These features enable GeneSys to nominate key genes governing state transitions in a developmental system, supporting both mechanistic insight and hypothesis generation. Together, GeneSys provides a flexible and extensible framework to denoise single-cell data and simulate transcriptomic developmental landscape guided by known or hypothesized developmental constraints, empowering the discovery of regulatory mechanisms from high-dimensional single-cell datasets.